AUDIO SIGNAL FILTERING OF INDUSTRIAL EQUIPMENT BASED ON AUTOENCODER

Abstract

The article considers the actual problem of filtering noise in audio signals generated by industrial equipment in order to improve the efficiency of modern automated control and management systems. The different approaches to filtering the noise components of audio signals are considered, with special attention focused on auto-encoders based on fully connected and convolutional neural networks. A series of experiments were conducted to analyze the effect of the autoencoder architecture parameters on the quality of noise filtering, in particular, the effect of the neural network “bottleneck” size on the overall system performance was also studied. In the process of the study, artificially generated signals with different spectral characteristics that model the conditions of an industrial environment are used as noise sources. To evaluate the filtering efficiency, the signal-to-noise ratio (SNR) metric is used to estimate the quality of the target signal recovery, which is the audio signal of an industrial engine. The experimental results demonstrate that both autoencoder architectures show a high ability to clean the noised signal. This study confirms that modern neural networks can significantly improve the quality of filtering, providing a reliable tool for monitoring the state of equipment in real time, which is especially important for systems focused on early detection of failures and prevention of accidents. The study also emphasizes the importance of using auto-encoders in monitoring tasks in terms of their adaptability to changes in the environment and their ability to self-learn. Due to the ability of auto-encoders to separate useful signals from noise interference, the system is able to provide high sensitivity to small changes in equipment operation, which is extremely important in industrial environments to ensure continuous operation. Auto-encoders also demonstrate the ability to scale within complex signal processing systems, making them suitable for use in enterprises with a large amount of equipment. In addition, their usage helps to reduce the need for professional assistance in the monitoring process, as neural network models are able to adapt to new conditions and learn from the changing data coming from the real environment. The practical significance of the study lies in its application to a wide range of industrial tasks related to improving equipment reliability and reducing maintenance costs. The use of auto-encoders in diagnostic systems can significantly reduce the risks of unpredictable failures, as well as optimize equipment maintenance costs through timely diagnostics and maintenance.